CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-11 10:36:54 +00:00
Code Issues Packages Projects Releases Wiki Activity

vulkan: optimize mul_mat for small values of N (#10991)

Browse Source 
Make the mul_mat_vec shaders support N>1 (as a spec constant, NUM_COLS) where
the batch_strides are overloaded to hold the row strides. Put the loads from the
B matrix in the innermost loop because it should cache better.

Share some code for reducing the result values to memory in mul_mat_vec_base.
This commit is contained in:
Jeff Bolz
2024-12-30 11:27:11 -06:00
committed by GitHub
parent c250ecb315
commit 716bd6dec3
9 changed files with 288 additions and 349 deletions
Download Patch File Download Diff File Expand all files Collapse all files

92
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp
View File

@@ -5,11 +5,6 @@
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
layout (constant_id = 0) const uint BLOCK_SIZE = 32;
layout (constant_id = 1) const uint NUM_ROWS = 1;
shared FLOAT_TYPE tmpsh[NUM_ROWS][BLOCK_SIZE];
void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
uint a_offset, b_offset, d_offset;
get_offsets(a_offset, b_offset, d_offset);
@@ -32,24 +27,17 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
const uint s_offset = 8*v_im;
const uint y_offset = 128*v_im + l0;
FLOAT_TYPE temp[NUM_ROWS];
FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
[[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
temp[i] = FLOAT_TYPE(0);
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
[[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
temp[j][i] = FLOAT_TYPE(0);
}
}
[[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
const uint y_idx = i * QUANT_K + y_offset;
B_TYPE_VEC2 b0 = data_b_v2[(b_offset + y_idx) / 2 + 0];
B_TYPE_VEC2 b16 = data_b_v2[(b_offset + y_idx) / 2 + 8];
B_TYPE_VEC2 b32 = data_b_v2[(b_offset + y_idx) / 2 + 16];
B_TYPE_VEC2 b48 = data_b_v2[(b_offset + y_idx) / 2 + 24];
B_TYPE_VEC2 b64 = data_b_v2[(b_offset + y_idx) / 2 + 32];
B_TYPE_VEC2 b80 = data_b_v2[(b_offset + y_idx) / 2 + 40];
B_TYPE_VEC2 b96 = data_b_v2[(b_offset + y_idx) / 2 + 48];
B_TYPE_VEC2 b112 = data_b_v2[(b_offset + y_idx) / 2 + 56];
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
f16vec2 d = data_a[ib0 + i].d;
@@ -74,48 +62,42 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
uvec2 qs0 = uvec2(unpack8(qs0_u16));
uvec2 qs16 = uvec2(unpack8(qs16_u16));
FLOAT_TYPE sum1 = FLOAT_TYPE(0.0);
FLOAT_TYPE sum2 = FLOAT_TYPE(0.0);
[[unroll]] for (int l = 0; l < 2; ++l) {
sum1 = fma(FLOAT_TYPE(b0[l]), FLOAT_TYPE(s0_lo4[0]) * FLOAT_TYPE((qs0[l] >> 0) & 3),
fma(FLOAT_TYPE(b16[l]), FLOAT_TYPE(s0_lo4[1]) * FLOAT_TYPE((qs16[l] >> 0) & 3),
fma(FLOAT_TYPE(b32[l]), FLOAT_TYPE(s0_lo4[2]) * FLOAT_TYPE((qs0[l] >> 2) & 3),
fma(FLOAT_TYPE(b48[l]), FLOAT_TYPE(s0_lo4[3]) * FLOAT_TYPE((qs16[l] >> 2) & 3),
fma(FLOAT_TYPE(b64[l]), FLOAT_TYPE(s4_lo4[0]) * FLOAT_TYPE((qs0[l] >> 4) & 3),
fma(FLOAT_TYPE(b80[l]), FLOAT_TYPE(s4_lo4[1]) * FLOAT_TYPE((qs16[l] >> 4) & 3),
fma(FLOAT_TYPE(b96[l]), FLOAT_TYPE(s4_lo4[2]) * FLOAT_TYPE((qs0[l] >> 6) & 3),
fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_lo4[3]) * FLOAT_TYPE((qs16[l] >> 6) & 3), sum1))))))));
sum2 = fma(FLOAT_TYPE(b0[l]), FLOAT_TYPE(s0_hi4[0]),
fma(FLOAT_TYPE(b16[l]), FLOAT_TYPE(s0_hi4[1]),
fma(FLOAT_TYPE(b32[l]), FLOAT_TYPE(s0_hi4[2]),
fma(FLOAT_TYPE(b48[l]), FLOAT_TYPE(s0_hi4[3]),
fma(FLOAT_TYPE(b64[l]), FLOAT_TYPE(s4_hi4[0]),
fma(FLOAT_TYPE(b80[l]), FLOAT_TYPE(s4_hi4[1]),
fma(FLOAT_TYPE(b96[l]), FLOAT_TYPE(s4_hi4[2]),
fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_hi4[3]), sum2))))))));
[[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
B_TYPE_VEC2 b0 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 0];
B_TYPE_VEC2 b16 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 8];
B_TYPE_VEC2 b32 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 16];
B_TYPE_VEC2 b48 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 24];
B_TYPE_VEC2 b64 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 32];
B_TYPE_VEC2 b80 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 40];
B_TYPE_VEC2 b96 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 48];
B_TYPE_VEC2 b112 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 56];
FLOAT_TYPE sum1 = FLOAT_TYPE(0.0);
FLOAT_TYPE sum2 = FLOAT_TYPE(0.0);
[[unroll]] for (int l = 0; l < 2; ++l) {
sum1 = fma(FLOAT_TYPE(b0[l]), FLOAT_TYPE(s0_lo4[0]) * FLOAT_TYPE((qs0[l] >> 0) & 3),
fma(FLOAT_TYPE(b16[l]), FLOAT_TYPE(s0_lo4[1]) * FLOAT_TYPE((qs16[l] >> 0) & 3),
fma(FLOAT_TYPE(b32[l]), FLOAT_TYPE(s0_lo4[2]) * FLOAT_TYPE((qs0[l] >> 2) & 3),
fma(FLOAT_TYPE(b48[l]), FLOAT_TYPE(s0_lo4[3]) * FLOAT_TYPE((qs16[l] >> 2) & 3),
fma(FLOAT_TYPE(b64[l]), FLOAT_TYPE(s4_lo4[0]) * FLOAT_TYPE((qs0[l] >> 4) & 3),
fma(FLOAT_TYPE(b80[l]), FLOAT_TYPE(s4_lo4[1]) * FLOAT_TYPE((qs16[l] >> 4) & 3),
fma(FLOAT_TYPE(b96[l]), FLOAT_TYPE(s4_lo4[2]) * FLOAT_TYPE((qs0[l] >> 6) & 3),
fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_lo4[3]) * FLOAT_TYPE((qs16[l] >> 6) & 3), sum1))))))));
sum2 = fma(FLOAT_TYPE(b0[l]), FLOAT_TYPE(s0_hi4[0]),
fma(FLOAT_TYPE(b16[l]), FLOAT_TYPE(s0_hi4[1]),
fma(FLOAT_TYPE(b32[l]), FLOAT_TYPE(s0_hi4[2]),
fma(FLOAT_TYPE(b48[l]), FLOAT_TYPE(s0_hi4[3]),
fma(FLOAT_TYPE(b64[l]), FLOAT_TYPE(s4_hi4[0]),
fma(FLOAT_TYPE(b80[l]), FLOAT_TYPE(s4_hi4[1]),
fma(FLOAT_TYPE(b96[l]), FLOAT_TYPE(s4_hi4[2]),
fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_hi4[3]), sum2))))))));
}
temp[j][n] = fma(dall, sum1, fma(-dmin, sum2, temp[j][n]));
}
temp[n] = fma(dall, sum1, fma(-dmin, sum2, temp[n]));
}
}
// sum up partial sums and write back result
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
tmpsh[n][tid] = temp[n];
}
barrier();
[[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) {
if (tid < s) {
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
tmpsh[n][tid] += tmpsh[n][tid + s];
}
}
barrier();
}
if (tid == 0) {
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
data_d[d_offset + first_row + n] = D_TYPE(tmpsh[n][0]);
}
}
reduce_result(temp, d_offset, first_row, num_rows, tid);
}
void main() {